DURIP GPU Cluster to Enable Physics-Informed Machine Learning for Underwater Robot Design
Abstract
Ongoing MURI Project. The MURI project entitled, #SERPENT: Self EneRgetic Propulsion ENTity,# aims to design, build, and deploy a long-endurance underwater robot. The four research thrusts center around (1) efficient swimming and control of the SERPENT robot, (2)the electrolyte used for both energy storage and hydraulic actuation, (3) mechanisms for wave energy harvesting, and (4) maximizingendurance through system design optimization of the SERPENT robot.Proposed DURIP Project. This DURIP project will support the procurement of state-of-the-art high-performance computing (HPC) graphical processing units (GPUs). It will enhance the aforementioned MURI project by improving the accuracy of models and accelerating critical computations involved in developing the SERPENT robot usingGPUs. Specifically, there are four project objectives for this DURIP. The first objective of this DURIP project is to apply physicsinformed machine learning as an approach for combining physics-based modelsand experimental data of the hydrodynamics and mechanicsof SERPENT, in order to improve the accuracy of the models used for design optimization. The second objective is to reduce the computation time for the aforementioned physics-based models, especially when solving partial differential equations on fine meshes thatevolve over time. These first two DURIP objectives enhance the research capabilities for Thrust 4 of the SERPENT project. The thirdobjective of this DURIP project is to accelerate the construction of fast and approximate models of SERPENT, which will be used forreal-time control. The fourth and final objective is to accelerate algorithms involved in imitation learning of the swimming motionfor SERPENT. The third and fourth DURIP objectives enhance the research capabilities for Thrust 1 of the SERPENT project.We proposeto purchase two GPU nodes (NVIDIA Ampere A100 GPUs) and collaborate with UC San Diego#s Triton Shared Computing Cluster (TSCC) to manage the computinghardware. The 8 GPUs and 128 CPU cores on these nodes (total) will accelerate the training and evaluation of neural networks and enable efficient matrix computations inherent in physics-based models. These nodes will have a combined 2 TB of memory to support the large memory needs for backpropagation in the neural networks and sensitivity analysis in the physics-based models. Partnering with TSCC provides numerous advantages, including the management of computing systems by high-performance computing experts, which results in minimal operational overhead. Moreover, TSCC offers pre-installed software packages for running physics-basedsimulations, enabling us to begin research immediately. The anticipated budget is $131,445. This includes the two GPU nodes ($127,193) and installation costs ($4252).
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Mar 15, 2024
- Source ID
- N000142412222
Entities
People
- John Hwang
Organizations
- Office of Naval Research
- United States Navy
- University of California, San Diego